CATEGORY · 17 TOOLS

This category covers the building blocks behind AI products: model APIs such as Google AI Studio and Cohere, routers such as OpenRouter, open-weight model families from DeepSeek, Meta, Alibaba and Mistral, local runtimes such as Ollama and LM Studio, and supporting infrastructure for vector search, evaluation, notebooks and GPU compute.

WHERE THESE 17 TOOLS KEEP DATA

  • AU hosting option3
  • Can run onshore8
  • Runs on your device2
  • No AU option published4

001 / OUR PICKS

Two trends defined 2025 and 2026. Open-weight models narrowed the gap with closed models for many business tasks, making it practical to run capable models on your own hardware or cloud account. At the same time, the Model Context Protocol became a common way to connect models to tools and data, so frameworks and platforms now compete on how well they support agents rather than plain chat.

Australian users include software companies embedding AI in their products, data teams in banks, universities and government building internal assistants, and agencies delivering custom systems for clients. Many choose their stack around one requirement first: whether prompts and data must stay in an Australian region.

ALSO IN THIS CATEGORY · 11

002 / HOW TO CHOOSE

  1. 01

    Hosted API, cloud platform or self-hosted

    Direct vendor APIs are quickest to start. Cloud platforms such as AWS Bedrock, Azure and Google Cloud add enterprise contracts and regional hosting. Self-hosting open models gives the most control but needs GPUs, monitoring and staff who can keep it running.

  2. 02

    Licence terms for open models

    Open-weight does not always mean open source. Read each model's licence for commercial use limits, attribution rules and acceptable use restrictions, and check whether fine-tuned versions inherit them. Record which licence applies to every model in production.

  3. 03

    Region and provider due diligence

    Check which regions a model is offered in, and who operates the endpoint. A model served by a third-party router or inference host may run in a different country from the original developer. Government and critical infrastructure work may also restrict certain vendors entirely.

  4. 04

    Observability and evaluation

    Production systems need tracing, cost tracking and automated tests that compare model outputs over time. Choose tooling that stores traces where you are comfortable holding customer messages, and that lets you swap models without rewriting your application.

  5. 05

    Cost controls

    Usage-based pricing can grow quickly with agents and long documents. Set spending limits per key, cache repeated prompts, route simple tasks to smaller models and monitor token use per feature so you can see which parts of your product are expensive.

003 / FOR AUSTRALIAN BUSINESSES

Data that can't leave Australia? →

AWS, Microsoft Azure and Google Cloud all operate Sydney and Melbourne regions, and several offer selected frontier and open models there, which is the usual route for keeping inference onshore under the Privacy Act 1988 and client contracts. Direct vendor APIs and routers generally process requests in the US. Australian Government entities must block DeepSeek products under PSPF Direction 001-2025, and suppliers to government should expect similar questions. Training or fine-tuning on customer data needs a clear collection notice and purpose. Local runtimes such as Ollama keep everything on your own machine. Cloud marketplaces can bill in Australian dollars through an existing account, while most AI-native vendors bill in US dollars.

004 / QUESTIONS

Anything else is quicker to answer on a call. Thirty minutes, free, with the people who would build it.

Use models offered in an Australian region through AWS Bedrock, Azure or Google Cloud, or run open-weight models on your own servers or GPU instances in Sydney or Melbourne. Check that logging, vector databases and evaluation tools are also onshore, because traces and embeddings often contain the same personal information as the prompts.

Thirty minutes, free. Tell us the job and we'll tell you which tool we'd use, and how we'd set it up.

Book a free call→